In context
In early 2024, visual object tracking remained a critical enabler for robotics applications such as human-robot interaction, intelligent transportation, and autonomous driving. Most state-of-the-art Siamese trackers relied on either cross-correlation or Transformer-based feature fusion, but rarely leveraged both, often losing semantic information or falling into local optima. This paper from Chinese robotics researchers addressed that gap by proposing a dual fusion architecture.
What was reported
Researchers from Shenyang University of Chemical Technology, the Chinese Academy of Sciences' State Key Laboratory of Robotics, and collaborating institutions introduced CTTrack, a single-object tracking algorithm that fuses template and search-region features through two parallel branches: an improved cross-correlation layer and a Transformer layer. The design aims to combine local correlation strengths with global attention modeling.
The backbone is a modified Swin-Transformer (stage 3 output, stride 16, 384 channels). The cross-correlation branch incorporates a similarity matrix to enhance features related to the target in the current frame, improving matching accuracy. The Transformer branch uses multi-head self- and cross-attention to model global dependencies. Outputs from both branches are concatenated and refined with depthwise convolution.
On five public benchmarks, the algorithm achieved success rates of 81.8% on TrackingNet, 65.7% on LaSOT, 66.2% on NFS, 69.4% on UAV123, and 69.8% on OTB2015, with an average tracking speed of 40 frames per second.
Why it mattered
By demonstrating that complementary fusion of local and global feature interactions can improve both accuracy and robustness without sacrificing real-time speed, this work offered a practical architecture for vision systems in industrial robotics and autonomous platforms, where reliable object tracking under varying conditions is essential.
“The advantages of these two fusion methods are complementary, so that template and search area features can fully interact.”
Source: 《机器人》期刊 (robot.sia.cn) · Published 2024-02-04 · “基于互相关－Transformer双层特征融合的目标跟踪算法”